OpenSensations™Models
OpenSensations curates and develops open models that read sensations in text. Each one is published with open weights on Hugging Face, so you can download it and run it yourself, or call it through the API.
There are 2 today: a small one and a larger one, fine-tuned for the same task so they can be compared on the same text. The site and the API are answered by the small one. Read the limitations before using its scores.
OpenSensations RoBERTa
Small and fast. The default.
| Hugging Face | mts-the-architect/open-sensations-roberta-125m-fp32 |
|---|---|
| Base model | FacebookAI/roberta-base |
| Size | 125 million parameters |
| How it was made | Full fine-tune |
| Licence | MIT |
Trained on the Lancaster Sensorimotor Norms (about 40,000 words) and the contextualized sensorimotor norms (448 sentences for 112 words). On words it never saw, its scores correlate 0.81 with people's ratings.
OpenSensations Gemma
A larger model, for comparison on the same text.
| Hugging Face | mts-the-architect/open-sensations-gemma4-e2b-bf16-lora |
|---|---|
| Base model | google/gemma-4-E2B |
| Size | Gemma 4 E2B with a small adapter |
| How it was made | Adapter (LoRA) and classifier head |
| Licence | Apache-2.0 for the adapter. The base model has its own terms. |
Trained on the contextualized sensorimotor norms only (448 sentences for 112 words): it is still the first prototype, and isn't served here yet.
The download is the adapter and the classifier head, not the base model. It runs with the loader in its repository, not a generic text pipeline.
What they score
Both models learned from sensorimotor norms: people's ratings, from 0 to 5, of how strongly a word is tied to each sense. Both score the same 6 labels, each from 0 to 1 and independently of the others, so several can be high at once.
vision, hearing, smell, taste, touch, interoception
Input: a target word and the sentence it is in. For a whole sentence, the site scores every word that carries meaning as the target and keeps each sense's highest score.
Choosing a model
A classify request can name the model it wants in its implementation field; without one, the default model answers. GET /v1/meta lists the models on offer. Every result names the model and the exact revision that produced it, so a score can always be traced to the weights behind it.
How well they read
Each model's card on Hugging Face reports its results on a held-out test set, label by label, with the training settings and the data it used. Read the limitations before relying on a score.
Contribute a model
Others can contribute a fine-tuned model for this task. The rules are on AI Attributes™.